Meta launched Muse Spark 1.1, an agentic model that delegates to subagents, does computer use, and actively manages a 1M-token window, served by the new Meta Model API in public preview and compatible with OpenAI's API: the open-weights company now offers a hosted model.
On July 9, 2026, Meta Superintelligence Labs unveiled Muse Spark 1.1, a multimodal reasoning model geared toward agentic tasks, and made it available to developers through a quiet but significant new offering: the Meta Model API, now in public preview. Meta — the company that built its reputation giving away the Llama weights so anyone could run them wherever they wanted — is now starting to offer a model served from its own infrastructure too. It's not a 180-degree reversal — open weights are still part of its DNA — but it is a move into the hosted-service territory that OpenAI and Anthropic dominate.
This isn't a model that answers and goes quiet, it's one built to do the work
Muse Spark 1.1 is designed to act, not just to chat. It runs tools, does tool calling across multiple applications, and operates in computer use flows — meaning it drives an interface the way a person would in front of a screen. And it doesn't work alone: it orchestrates subagents that split up and carry out parts of a task, with the main model integrating the results afterward. It's the difference between an employee who does everything themselves and a team lead who delegates and reassembles. For anyone building agents, that native multi-agent orchestration capability matters more than any single-response benchmark.
Memory as a first-class capability
The most interesting technical detail isn't the size of the context window, but how it manages it. Muse Spark 1.1 has a 1 million token window, but instead of dragging the whole thing around like dead weight, it actively manages it. In Meta's own words, it 'remembers actions, retrieves information from work far earlier, and compacts in a way that preserves the critical steps needed for later work.' Compacting while preserving what matters, rather than blindly truncating when the window fills up, is exactly the bottleneck that breaks long-horizon agents: the one that forgets step three by the time it reaches step thirty fails. Having the model treat memory as a capability rather than a passive limit is what makes it fit for tasks that run for hours, not seconds.
The move that rearranges the board: speaking your rival's language
Here's the play that almost goes unnoticed next to the model itself. The Meta Model API is presented as compatible with OpenAI's API — in the words of Replit's CEO quoted by Meta, 'a clean, OpenAI-compatible package.' In plain terms: much of the code already written against OpenAI can point to Meta's endpoint with minimal changes, without rewriting the integration from scratch. Meta is accepting, without saying it in those words, that 'how you talk to the model' is ceasing to be a proprietary asset and is becoming a de facto standard. When the open-weights provider also adopts its competitor's interface, interoperability stops being a community promise and becomes a business decision.
The lesson for anyone building with AI
The takeaway is ripe for the picking, and it holds up even against the data Meta hasn't published yet — pricing, credits, regional availability: when APIs converge toward a common format, loyalty to a single model evaporates and switching engines tends to look more like a config change than a three-month engineering project. Anyone chaining their product to a single provider with custom integrations is building a cage they pay for themselves. At NeuralOS, that conviction is design doctrine, not prediction: the platform is model-agnostic and the model is chosen by configuration, without rewriting the logic of your apps, agents, or automations. Muse Spark 1.1 doesn't change that bet; it reinforces the trend behind it. When even Meta speaks its rivals' language, the useful question stops being 'which model do I use' and becomes 'what do I build on top that survives the next model change.'
(Editor's note: as of this writing, Meta's official announcement describes the model, its agentic nature, the 1M-token window, and OpenAI compatibility, but does not publish pricing, trial credits, or details on regional availability. We'll update when official figures appear.)